Experiment and Problem

class optexp.Experiment(optim: Optimizer, problem: Problem, eval_every: int, steps: int, seed: int = 0, hardware_config: HardwareConfig = StrictManualConfig(num_devices=1, micro_batch_size=None, eval_micro_batch_size=None, num_workers=0, device='auto'), group: str = 'default')

Specify an experiment.

Parameters:

• optim (Optimizer) – optimizer to use.

• problem (Problem) – problem to solve.

• eval_every (int) – often to evaluate the metrics.

• steps (int) – total number of steps. To convert from epochs, use optexp.utils.epochs_to_steps().

• seed (int, optional) – seed for the random number generator. Defaults to 0.

• hardware_config (HardwareConfig, optional) – implementation details. Defaults to StrictManualConfig().

• group (str, optional) – name for logging. Defaults to "default".

class optexp.Problem(model: Model, dataset: Dataset, batch_size: int, lossfunc: LossLikeMetric, metrics: Iterable[Metric], datapipe: DataPipe = TensorDataPipe(), init_callback: InitCallback | None = None)

Specify a problem.

Parameters:

• model (Model) – model to optimize.

• dataset (Dataset) – dataset to fit the model to.

• batch_size (int) – effective batch size. To use gradient accumulation, set the micro_batch_size in optexp.hardwareconfig.HardwareConfig.

• lossfunc (Metric) – loss function to use for optimization.

• metrics (Iterable[Metric]) – metrics to evaluate.